Hybrid Modeling of Arima, ANN and SVM for Macro Variables Forecasting in Pakistan

Abstract

Time series forecasting remains a challenging task owing to its nonlinear, complex and chaotic behavior. The purpose of this paper is to analyze the forecast performance of different models for Pakistan’s macroeconomic variables such as inflation, exchange rate and stock return. In which linear Autoregressive integrated moving average (ARIMA) model and nonlinear models like artificial neural networks (ANN) and Support vector machines (SVM) are employed. Then a hybrid methodology is used which combines the linear ARIMA with nonlinear models of ANN and SVM. The forecasting performance of all models i.e., ARIMA, ANN, SVM, ARIMA-ANN and ARIMA-SVM are compared on the basis of RMSE and MAE. The results indicate that the best forecasting model to achieve high forecast accuracy is the hybrid ARIMA-SVM.